Labs

GDPR-compliant AI for public administration

Data sovereignty, EU AI Act, national data protection laws — all covered. We build AI systems for authorities that run entirely on German servers. Open-source models, fully auditable, legally compliant.

What we build for authorities

Case Worker Assistant

Searches through specialist laws, internal regulations and case law in seconds. Provides answers with source references — case workers retain control and save research time.

Document Processing

Classify applications, extract form data, structure information. The AI reads invoices, contracts and official notices — faster and more accurate than manual processing.

Citizen Chatbot

Answers standard questions 24/7: opening hours, applications, responsibilities. Relieves the telephone switchboard and citizen services — without data ever leaving the town hall.

File & Record Summarization

Summarize meeting minutes, lengthy procedures and extensive files in seconds. With source references to the relevant passages for review.

How the project runs

1

Workshop & Analysis

We analyze your processes, identify use cases with the greatest impact and assess the legal framework. Output: a prioritized roadmap.

2

Pilot Project

In 8-12 weeks we build a first system with clearly defined scope. Typically: one concrete use case, measurable results, full documentation.

3

Rollout & Operations

Scaling to other departments or use cases. We handle monitoring, updates and training — or transfer the know-how to your team.

The problem: standard AI can't enter public offices

ChatGPT, Gemini, Claude — the well-known AI tools process data in the US. For German authorities this is taboo under GDPR. At the same time the pressure is rising: staff shortages, digitization mandates, EU AI Act. The solution: AI that runs exclusively on German servers, with open-source models and full auditability.

What AI actually does in a public administration today

Not a vision - use cases that already run in town halls and district offices, and what each of them requires.

Reviewing tender documents

The AI reads bids and specifications, checks completeness against the eligibility criteria, flags deviations and produces a tabular pre-evaluation of the bidders. The award decision stays with the procurement office - the review takes hours instead of days. Requirement: the documents stay in-house, so on-premise or EU cloud.

File research with citations

Case workers ask in plain language about cases, bylaws or internal directives and get the answer with a reference to the document and page. Technically this is a RAG system: the model may only answer from your documents and does not invent anything.

Triaging citizen requests

Incoming e-mails and forms are sorted by concern, responsibility and urgency, and standard cases get a draft reply right away. The inbox becomes a task list - and citizens no longer wait three weeks for "we will forward this".

Drafting notices and letters

From the file and your text modules, a draft emerges in the right tone, including the legal basis. Approval and responsibility stay with the case worker - under the EU AI Act, that is exactly what decides the risk class.

Meetings and minutes

Recordings of council and committee meetings become draft minutes with a list of resolutions and open items. Transcription runs locally; nothing goes to an external service.

With kommun.ai we run our own AI citizen assistant for municipalities - EU-hosted, multilingual, dialect-aware. That is how we know where the hurdles lie in day-to-day administration: rarely in the model, almost always in data quality, permissions and acceptance in the team.

Three operating models - and when each fits

On-premise in your own data centre

Maximum control, highest effort. Makes sense if you process personal data from social, youth or public order offices and run your own servers anyway. Needs GPU hardware and someone to maintain it.

At the municipal IT provider

The middle path for most municipalities: operation at the public IT provider that already hosts your specialist procedures. Contractually clean, IT baseline protection in place, no additional processor.

Sovereign EU cloud

Fastest start, good scaling. STACKIT, IONOS or Open Telekom Cloud with C5 attestation and processing exclusively in the EU. Usually the pragmatic route for pilots - and can be moved to one of the other two models later.

Legally airtight

Our solutions meet all relevant compliance requirements — audited and documented.

  • GDPR — Data processing exclusively within the EU, no third-country transfers
  • National data protection laws (BDSG) and specialist regulations
  • EU AI Act — Transparency, risk classification, documentation
  • IT baseline protection — compatible with BSI standards, audit logs
  • BITV 2.0 — Accessibility in all interfaces

Technology stack for sovereignty

We exclusively use open-source models and European infrastructure. No dependencies on US providers.

Models: Llama 3, Mistral, Teuken-7B (German model), Mixtral

Hosting: STACKIT (Schwarz Group), plusserver, Open Telekom Cloud, Hetzner, or on-premise

Infrastructure: Kubernetes, vector databases (Qdrant, Weaviate) — completely on-premise

Frequently Asked Questions

Bring AI into your office — securely

We provide non-binding consulting on your use cases and show you concretely what's feasible. Including an initial assessment of costs and timeline.

Schedule consultation
GDPR-Compliant AI for Government & Public Sector | CAZ Labs